Shape what your agents know

An agent's instructions and memories outlive its choice of model. You can keep working with the same teammate as models improve.

Write instructions

Open Agent settings and describe the agent's role, writing style, responsibilities, and when it should ask you for a decision. These are lasting instructions that you write; the agent cannot change them.

For example:

Help us prepare client updates. Use concise language, cite the sources behind factual claims, and ask us when a decision is needed.

Inspect memories

An agent can save short facts while working. Open Memories in its work details to see what it has learned. Correct a memory when something changes or remove information it should forget.

Memories belong to the agent and are shared across the workspace. Avoid saving information that other workspace members should not see.

Add context for the task

Attach files or enable a connected app when the agent needs source material. An attachment provides context for the work; it does not replace your lasting instructions.

Changing an agent's model keeps its identity, instructions, memories, and saved work. Different models may respond differently.